clinical practice guidelines, their representation
TRANSCRIPT
CLINICAL PRACTICE GUIDELINES, THEIR REPRESENTATION, ENACTMENT AND PROCESSING
Szymon WilkInstitute of Computing SciencePoznan University of TechnologyPoznan, Poland
INTRODUCTION
Clinical practice guidelines (CPGs)
• Introduced to limit the variations in service delivery and to minimize healthcare costs
• Initially aimed at nurses and other ancillary personnel, then adopted (slowly) by physicians
• Increasing popularity of computer-interpretable guidelines (CIGs) integrated with clinical systems
Knowledge-based tools for disease-specific patient management [Rosenfeld and Shiffman, 2009]
Diversified clinical terminology: guideline, algorithm, pathway…
CPGs (and CIGs) in clinical practice
• On the one hand, multiple advantages• Increased adoption of evidence-based medicine (EBM) and improved
adherence to standards of practice
• Positive impact on patient outcomes (e.g., decreased mortality)
• On the other hand, still limited adoption• Considered to be “cookbook medicine”
• Given mostly in paper format
• Limited standardization of formal representations
No support for multimorbid conditions
Clinical guidelines are only one option for improving the quality of care [Woolf et al., 1999]
Representation of CPGs and CIGs
• Text models (CPGs)• Limited to textual content (possibly long)
• Additional information (tags) augmenting the text
• Task-network models (CIGs)• Interpretable by a physician
• Aimed at (semi-) automatic analysis and enactment by a computer system
Leve
l of
form
aliz
atio
n
Sample CPGs and CIGs
Text models: GEM (Guideline Elements Model)
• A set of XML tags (100+) for marking selected parts of a guideline document
• Also a set of tools for viewing and tagging text documents
• Official standard endorsed by ASTM International, very popular, despite its simplicity
• Possibility of automatic discovery of clinical rules – requires more detailed tags (experimental extension)
http://gem.med.yale.edu
Task-network models
• A CPG represented as a directed graph with nodes capturing specific steps and arcs – dependencies between steps
• Basic types of• Data query – collecting data (from the user or other system)
• Decision – making a decision
• Action – performing some clinical action
• Invocation – invocation of another CIG
• Multiple representations based on the task-network model with varying complexity and focus (e.g., temporal aspects)
• Increasing popularity due to possible integration with HISs
Task-network model representations
• GLIF3 – a multiple-level representation with conceptual flowcharts and an executable specification
• SAGE – an complete (event-driven) environment for creating and executing CIGs, integration with HIS via vMR
• Asbru – a formalism (+ tools) focused on modeling temporal aspects and uncertainties associated with CIG execution
• PROforma – a formalism (+ tools) for creating and managing CIGs with focus on argumentation-based decision making, one of few commercial applications
• SDA* – a flowchart-based representation developed for the K4CARE project and aimed at team-based automatic execution
• BPMN – a model (+ tools) for representing business processes
PROforma and OpenClinical.net
• PROforma• A scientific project (Cancer Research UK, 1992-now)
• A commercial product Alium (Deontics, UK)
• OpenClinical.net• Web-based repository of CIGs represented in PROforma
• Peer-driven review and verification of submitted CIGs before they are made widely available
• Possibility of running (simulating) available algorithms through a web-based system
https://deontics.com/
Example – CIG for abdomen traumaSOR, chory z podejrzeniem
obrażeń brzucha
Stan
chorego?
Mechanizm
urazu?
Stan niestabilny
USG-FAST
Uraz nieprzenikający
Wynik
badania
obrazowego?
Poszukiwanie innego
źródła krwawienia
Wynik ujemny
Obserwacja i badanie
kontrolne
DPO
Wynik niepewny
Laparotomia
Wynik dodatni
Laparotomia
zwiadowcza
Uraz przenikający
Wynik
zwiadu?
Zwiad ujemny
Zwiad dodatni
Wynik
badania DPOWynik ujemny
Wynik dodatni
RTG / USG / TK
Stan stabilny lub niepewny
Wynik
badania
obrazowego?
Obserwacja objawów
otrzewnowych
Wynik ujemny
Objawy
otrzewnowe?
Brak objawów
otrzewnowych
Zaobserwowane objawy
otrzewnowe
Laparotomia
Wynik niepewny Wynik niepewny
1. Obserwacja objawów
otrzewnowych
2. Obserwacja stanu chorego
Stan
chorego?Zaobserwowane
objawy otrzewnowe
Bez zmian lub poprawa
1. Uzupełnienie krwi
2. Płynoterapia
3. Farmakoterapia
Pogarszający się
stan ogólny
1. Wywiad
2. Kontrola ABCDE
3. Badania szczegółowe
Definition of a decision step
A detailed specification of data items to be colleted
Specification of arguments for and againsta specific decision option
Supervised decision making
Provided patent data
Both decision options are possible
Supervised decision making
Highlighting arguments associated with specific decision options
Supervised decision making
A physician may overwrite any option suggested by a system (and select one with “poor” support).
If this option is selected, additional imaging procedures (defined in CIG)
will be prescribed.
Patterns of collaboration
• Goal-based workflow representation based on PROforma
• State-based exceptions for detecting obstacles and hazards and associated plans for handling them
• Formal description of two collaboration patterns• Assignment → provider is accountable for outcome and responsible for
handling exceptions
• Delegation → client is responsible for outcome and responsible for managing (selected) exceptions
M.A. Grando, M. Peleg, M. Cuggia, D. Glasspool: Patterns for Collaborative Work in Health Care Teams. Artificial Intelligence in Medicine, 2011, 53 (3), 139-160.
Patterns of collaboration
CIGs in BPMN
• Combination of BMPN (jBPM) and rules (Drools) to represent a CIG, integration with HIS via SOA
Rodriguez-Loya, S., Aziz, A., & Chatwin, C. (2014). A Service Oriented Approach for Guidelines-based Clinical Decision Support using BPMN. Studies in Health Technology and Informatics, 205, 43–47.
Decision rules associated with specific tasks to establish
decision outcomes
SUPPORT FOR MULTI-MORBIDITIES
Practical motivation
• 76% of people 65+ years old have 2+ chronic conditions, and their care costs are 5.5 times higher than for non multi-morbid patients [Bähler, et al. 2015]
Direct application of multiple CPGs “may have undesirable effects” and “diminish the quality of care” [Boyd et al. 2005]
No support for multi-morbid conditions is a “major shortcoming of CPG uptake in clinical practice” [Peleg, 2013]
Methodological challenge
“The challenge … to identify and eliminate redundant, contraindicated, potentially discordant, or mutually exclusive guideline based recommendations for patients presenting with comorbid conditions or multiple medications.” [Sittig et al., 2008]
One of the “grand challenges” for clinical decision support
A new, “combinatorial, logical, or semantic” methodological approach is needed [Fox et al. 2010]
LOGIC-BASED CIG PERSONALIZATION FRAMEWORK
Research goal and questions
1. How to represent rich primary (CPGs) and secondary (interactions, preferences) clinical knowledge?
2. What “reasoning” techniques to use to process knowledge encoded in the proposed formalism?
Goal: a framework for personalizing CPGs for multi-morbid patients by (1) mitigating adverse interactions and
(2) customizing resulting therapies based on patients’ preferences
Answer: first-order logic, theorem proving and model solving
Wilk, S., Michalowski, M., Michalowski, W., Rosu, D., Carrier, M., & Kezadri-Hamiaz, M. (2017). Comprehensive mitigation framework for concurrentapplication of multiple clinical practice guidelines. Journal of Biomedical Informatics, 66.
Patient preferences
• A new and important component of EBM
• Preferences are especially relevant when evidence is associated with a high level of uncertainty (→ “grey zone” or “preference-sensitive” decisions) [van der Weijden et al., 2013]
• Participation of patient groups already in the development of CPGs [van der Weijden et al., 2010]
EBM = Evidence + Experience + Preferences
“evidence is never enough to make treatment recommendations” [Murad et al., 2008]
First-order logic (FOL)
• A formal system to represent and reason about knowledge
• Knowledge represented using a domain-specific language with logical (fixed meaning) and non-logical symbols
• A theory 𝒟 is a collection of sentences
• An interpretation ℐ assigns the meaning (formal semantics) to non-logical symbols
• If ℐ satisfies all sentences in 𝒟, then it is called a model for 𝒟
• Theorem proving ( checking for consistency and entailment) and model finding techniques
FOL-based personalization framework
Primary knowledge (encoded in CPGs)
Secondary knowledge (not encoded in CPGs)
Possibly incomplete
Actionable graph 𝐴𝐺𝑑𝑖
• Captures a CPG for a given disease 𝑑𝑖• An intermediate representation based on a task-
network model for better interoperability
• Can be automatically obtained from other representations (e.g. GLIF3, SAGE)
• A directed graph with context, decision, action and parallel nodes (with additional attributes)
Mitigation-specific language 𝐿𝑚𝑖𝑡
• Allows describing all components of the mitigation problem
• Introduces structural and temporal predicates
Combined mitigation theory 𝒟𝑐𝑜𝑚𝑏
• Captures core components of a mitigation problem
• Defined as a triple 𝒟𝑐𝑜𝑚𝑚𝑜𝑛, 𝒟𝑐𝑝𝑔, 𝒟𝑝𝑖 , where
• 𝒟𝑐𝑜𝑚𝑚𝑜𝑛 – common axioms defining universal character of CPGs, e.g.,
• 𝒟𝑐𝑝𝑔 – a union of theories 𝒟𝑐𝑝𝑔𝑑1 ∪ 𝒟𝑐𝑝𝑔
𝑑2 ∪⋯∪ 𝒟𝑐𝑝𝑔𝑑𝑘 representing AGs
applied to a comorbid patient
• 𝒟𝑝𝑖 – a collection of available patient data (results of tests and
examinations, prescribed therapies, …)
Revision operator 𝑅𝑂𝑘
• Defines revisions to CPGs (𝒟𝑐𝑝𝑔) triggered by some undesired
circumstances (related to preferences or interactions)
• Defined as a pair 𝛼𝑘, 𝑂𝑝𝑘 , where
• 𝛼𝑘 – undesired circumstances that need to be addressed
• 𝑂𝑝𝑘 – a list of operations that revise 𝒟𝑐𝑝𝑔 (only) to address 𝛼𝑘
• Each 𝑂𝑝𝑘,𝑖 from 𝑂𝑝𝑘 defines a single find-and-replaceoperation ( replace, insert, delete)
Management scenario 𝒟𝑚𝑠
• Represents a safe (no interactions) and preferred (consistent with preferences) course of actions for a given patient
• Specifies clinical actions to be taken with their order and timing
• Introduces assumptions related to the future patient’s state
Mitigation Algorithm
Regular expression(RE) level
Revises 𝒟𝑐𝑝𝑔 to account for
patient preferences
Revises 𝒟𝑐𝑝𝑔 to avoid
interactions and establishes 𝒟𝑚𝑠
(depth-first search)
Revises 𝒟𝑐𝑝𝑔 according to the
currently applied 𝑅𝑂𝑘
CKD, AFib and HTN
CKD = chronic kidney disease, AFib = atrial fibrillation, HTN = hypertension
FOL representation –𝒟𝑐𝑝𝑔𝐴𝐹𝑖𝑏
exists x1,x2,x3,x4,x5,x6,x7,x8,x9,x10,x11,x12:(disease(x1,AFIB) /\ parallel(x2) /\ action(x7,BB) /\ parallel(x12)) /\(((disease(x1,AFIB) /\ parallel(x2) /\ decision(x3,AFIB_DUR) /\ result(x3,LT48H) /\ action(x4,FLEC)
/\ parallel(x5) /\ action(x8,CCB) /\ parallel(x12)) /\(disease(x1,AFIB) /\ parallel(x2) /\ decision(x3,AFIB_DUR) /\ result(x3,LT48H) /\ action(x4,FLEC)
/\ parallel(x5) /\ action(x9,ACEI) /\ parallel(x12))) \/ ((disease(x1,AFIB) /\ parallel(x2) /\ decision(x3,AFIB_DUR) /\ result(x3,GE48H)
/\ parallel(x5) /\ action(x8,CCB) /\ parallel(x12)) /\(disease(x1,AFIB) /\ parallel(x2) /\ decision(x3,AFIB_DUR) /\ result(x3,GE48H)
/\ parallel(x5) /\ action(x9,ACEI) /\ parallel(x12)))) /\((disease(x1,AFIB) /\ parallel(x2) /\ decision(x6,CHA2DS2) /\ result(x6,GE1)
/\ action(x10,WAR) /\ parallel(x12)) \/ (disease(x1,AFIB) /\ parallel(x2) /\ decision(x6,CHA2DS2) /\ result(x6,EQ0)
/\ action(x11,ASA) /\ parallel(x12)))
Simplified representation for brevity (e.g. no directPrec)
Interaction-Related Revision Operators
• 𝑅𝑂𝑖𝑛𝑡1 : if patient diagnosed with HTN and CKD, then remove
Step 1 from CPG for HTN
• 𝑅𝑂𝑖𝑛𝑡2 : if patient diagnosed with HTN, AFib and CKD, then
remove diuretics from CPG for HTN
• 𝑅𝑂𝑖𝑛𝑡3 : if patient diagnosed with AFib and CKD and anemia is
present, then replace DOAC with warfarin in CPG for AFib
• 𝑅𝑂𝑖𝑛𝑡4 : if patient diagnosed with AFib and CKD and GFR < 60,
then replace BB with metoprolol in CPG for AFib
Patient scenarioA 70-years old male with CKD and HTN having the following characteristics:(1) Decreased kidney function (GRF < 60) and mild anemia require ESA, patient does not
have metabolism abnormalities and is on CV risk and lifestyle management(2) HTN is managed according to Step 3 for uncontrolled BP
Patient scenario
Personalization framework is invoked
1. customize procedure applies 𝑅𝑂𝑝𝑟𝑒𝑓1 and revises CPG for AFib
2. mitigate procedure checks applicable 𝑅𝑂𝑖𝑛𝑡𝑘
• 𝑅𝑂𝑖𝑛𝑡1 → changes affect past actions (step 1 in CPG for HTN) and thus they are not
introduced by revise procedure
• 𝑅𝑂𝑖𝑛𝑡2 → diuretics are removed from CPG for HTN
• 𝑅𝑂𝑖𝑛𝑡3 → apixaban is discarded and warfarin is restored in CPG for AFib
• 𝑅𝑂𝑖𝑛𝑡4 → BB in replaced by metoprolol in CPG for AFib
For the last 12 hours patient has been experiencing irregular pulse, breathlessness, dizziness, and chest discomfort. Upon admission to the ED patient has been diagnosed with AFib that has been confirmed by standard ECG recording. Patient’s CHA2DS2 score is 2.
Patient has expressed preferences related to AFib therapy:
𝑅𝑂𝑝𝑟𝑒𝑓1 : if diagnosed with AFib and prescribed warfarin, then replace warfarin with apixaban
(one of the DOACs)
Revised CPGs
Marked elements constitute the management scenario 𝒟𝑚𝑠
Implementation and extension
• Complex FOL-based representation of primary and secondary knowledge, but hidden from clinicians
• Transforming the reasoning problem into planning one expressed in the Planning Domain Definition Language (PDDL)
GOAL-BASED MITIGATION FRAMEWORK
Goal-based mitigation framework
• Employs knowledge in National Drug File – Reference Terminology (NDF-RT) with information about prevention, treatment and physiological effects
• Relies on PROforma, however, CIGs are represented as high-level plans associated with goals with statements from NDF-RT
• Use of SNOMED-CT to encode information and HL7 FHIR to exchange information between PROforma engine and HIS
• Controller component oversees all events associated with CIG enactment, identifies conflicts and interact with a physician to solve these conflicts
Kogan, A., Tu, S. W., & Peleg, M. (2018). Goal-driven management of interacting clinical guidelines for multimorbidity patients. AMIA ... Annual SymposiumProceedings. AMIA Symposium, 2018(October), 690–699.
NDF-RT
Physiological Effect fulfills the abstract goal“prevent blood clots”
DU is-a Hemorrhagic Disorder
Patient ScenarioA patient who had cardiovascular disease (CVD) and, following the CIG’s recommendations for secondary prevention of CVD via decreased platelet aggregation, was started on aspirin. The patient developed duodenal ulcer (DU).
Goal forest
Controller retrieves from PROfoma and controls a current goal forest
Controller rules
Rules for controlling the forest, indicating conflicts and solving them (by interacting with the user)
Sequence diagram of interactions
Control rules (d) and (e)
Control rule(f)